Building AI agents with Claude in Google Cloud's Vertex AI

Key Concepts

• AI Agents: Software entities designed to perform tasks autonomously, often leveraging Large Language Models (LLMs).
• Agent Development Kit (ADK): An open-source, code-first framework for building, evaluating, and deploying AI agents.
• Multi-Cloud Playground (MCP): A standard protocol for LLMs and agents to access context and tools.
• Vertex AI Agent Engine: A managed platform for deploying, managing, and scaling AI agents in production.
• Agent-to-Agent Protocol: An open-source protocol enabling seamless communication and collaboration between agents built with different frameworks.
• LLMs (Large Language Models): Powerful AI models used as the "brain" of agents. Examples include Claude.
• Tools: Mechanisms to extend agent capabilities, providing access to specific skills or functionalities (e.g., calendar scheduling).
• Runners: Components that manage agent execution, including session state and interaction via CLI or web UI.
• Sessions: Mechanisms to store conversation history and maintain context across interactions with an agent.
• Agent Skills: Describe the functions or capabilities of an agent.
• Agent Card: A digital business card for an agent, allowing other agents or applications to understand its skills and how to interact with it.
• Agent Executor: Manages communication, requests, and responses between agents.

Challenges in Productionalizing AI Agents

• Fragmented Landscape: Numerous frameworks and tools exist, making integration complex.
• Interoperability Issues: Difficulty in enabling communication between agents built with different frameworks.
• Operational Overhead: Managing agents in production requires significant monitoring, logging, and governance efforts.

Google Cloud's Agent Stack

Google Cloud addresses these challenges with a four-component agent stack:

1. Agent Development Kit (ADK):
   • Open-source framework for building, evaluating, and deploying agents.
   • Code-first and developer-friendly.
2. Multi-Cloud Playground (MCP) Compatibility:
   • Standardizes agent communication with tools and context.
   • Allows agents to consume tools and connect with other agents seamlessly.
3. Vertex AI Agent Engine:
   • Managed platform for deploying, managing, and scaling AI agents in production.
   • Handles operational challenges and provides necessary capabilities.
4. Agent-to-Agent Protocol:
   • Open-source protocol for seamless communication and collaboration between agents, regardless of the framework used to build them.

Accessing Cloud Models on Vertex AI

• Vertex AI Model Garden: A centralized hub for discovering, deploying, and managing foundational and open models, including Cloud.
• Vertex AI Studio: A prompt UI for testing models.
• Cloud models are accessible through API or the console after providing credentials.

Building a Simple Agent with ADK: Birthday Planner

• Core Concepts: LLM agent, tools, runner, session.
• Process:
  1. Import necessary classes (LLM agent, Cloud).
  2. Define the agent using the LLM agent class:
     • Set the model (e.g., Cloud 3.7 Sonnet).
     • Give the agent a name.
     • Provide a description and instructions.
  3. Interact with the agent using the ADK CLI (ADK run <agent_name>) or web UI.
• Files Required: agent.py (agent logic), .env (environment variables), __init__.py.
• Cloud Integration: Supported through LLM registry.

Extending Agents with MCP: Multi-Agent System

• Scenario: Enhancing the birthday planner agent with a calendar service agent for scheduling.
• Components: Birthday planner agent, calendar service agent (using MCP), orchestrator agent.
• MCP Integration:
  • Use existing MCP servers as tools within ADK.
  • Deploy ADK-built tools using MCP for interaction with other agents.
• Orchestrator: Routes requests to the appropriate agent based on the user's intent.
• Web UI: Provides a way to debug and interact with the multi-agent system, showing which agent is being used for each task.

Deploying Agents on Vertex AI Agent Engine

• Challenges of Manual Deployment: Requires wrapping agent code in services like Fast API, building containers, managing infrastructure, and handling operations.
• Agent Engine Benefits: Simplifies deployment with a command like agent engine create.
• Features:
  • Observability capabilities and monitoring.
  • Automatic collection of interaction logs for evaluation.
  • Integration with Vertex AI Evaluation Service.
  • Support for agents built with various frameworks (ADK, LangGraph, LlamaIndex).
• Process:
  1. Provide base requirements for the agent.
  2. Use the provided class to create an agent endpoint on Agent Engine.
  3. Monitor the deployment in the Vertex AI console.

Agent-to-Agent Protocol (Bonus)

• Problem: Lack of a standardized way to connect agents built with different frameworks.
• Solution: Agent-to-Agent Protocol, an open protocol designed to foster agent collaboration.
• Key Concepts:
  • Agent Skills: Describe the functions or capabilities of the agents.
  • Agent Card: A digital business card for the agent, allowing other agents or applications to understand its skills and how to interact with it.
  • Agent Executor: Manages communication, requests, and responses between agents.
• Benefits: Enables building complex systems with agents communicating and collaborating to achieve tasks.
• Enterprise-Ready: Designed with features for governance and security.

Conclusion

The presentation outlines Google Cloud's agent stack, comprising ADK, MCP compatibility, Agent Engine, and Agent-to-Agent Protocol, as a solution to the challenges of building and deploying AI agents in production. It provides a practical demonstration of building a simple agent, extending it with MCP, deploying it on Agent Engine, and introduces the concept of Agent-to-Agent Protocol for enabling communication between agents built with different frameworks. The key takeaway is that Google Cloud offers a comprehensive set of tools and services to standardize agent development, simplify deployment, and facilitate collaboration, ultimately enabling the creation of powerful and scalable AI agent systems.